Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “model inference”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 1,549 records · Page 86

Variable protostellar accretion with episodic bursts

We present the latest development of the disk gravitational instability and fragmentation model, originally introduced by us to explain episodic accretion bursts in the early stages of star formation. Using our numerical hydrodynamics model with improved disk thermal balance and star-disk interaction, we computed the evolution of protostellar disks formed from the gravitational collapse of prestellar cores. In agreement with our previous studies, we find that cores of higher initial mass and angular momentum produce disks that are more favorable to gravitational instability and fragmentation, while a higher background irradiation and magnetic fields moderate the disk tendency to fragment. The protostellar accretion in our models is time-variable, thanks to the nonlinear interaction between different spiral modes in the gravitationally unstable disk, and can undergo episodic bursts when fragments migrate onto the star owing to the gravitational interaction with other fragments or spiral arms. Most bursts occur in the partly embedded Class I phase, with a smaller fraction taking place in the deeply embedded Class 0 phase and a few possible bursts in the optically visible Class II phase. The average burst duration and mean luminosity are found to be in good agreement with those inferred from observations of FUors. The model predicts the existence of two types of bursts: the isolated ones, showing well-defined luminosity peaks separated with prolonged periods (∼10{sup 4} yr) of quiescent accretion, and clustered ones, demonstrating several bursts occurring one after another during just a few hundred years. Finally, we estimate that 40%–70% of the star-forming cores can display bursts after forming a star-disk system.

79 ASTRONOMY AND ASTROPHYSICS↗

The Pantheon+ Analysis: Forward Modeling the Dust and Intrinsic Color Distributions of Type Ia Supernovae, and Quantifying Their Impact on Cosmological Inferences

Abstract Recent studies have shown that the observed color distributions of Type Ia supernovae (SNe Ia) can be well described by a combination of a dust distribution and an intrinsic color distribution. Using the Pantheon+ sample of 1701 SN Ia, we apply a new forward-modeling fitting method (Dust2Dust) to measure the parent dust and color distributions, including their dependence on host-galaxy mass. At each fit step, the SN Ia selection efficiency is determined from a large simulated sample that is reweighted to reflect the proposed distributions. We use five separate metrics to describe the goodness of fit: distribution of fitted light-curve color c , cosmological residual trends with c , cosmological residual scatter with c , fitted color–luminosity relationship β SALT2 , and intrinsic scatter σ int . We present the results and the uncertainty in 12-dimensional space. Furthermore, we measure that the uncertainty on this modeling propagates to an upper threshold uncertainty in the equation of state of dark energy w of 0.014(1) for the Pantheon+ cosmology analysis and contributes negligible uncertainty to the Hubble constant H 0 . The Dust2Dust code is made publicly available at https://github.com/djbrout/dustdriver .

79 ASTRONOMY AND ASTROPHYSICS↗

Emulator-based Bayesian calibration of a subglacial drainage model

Subglacial drainage models, often motivated by the relationship between hydrology and ice flow, sensitively depend on numerous unconstrained parameters. We explore using borehole water-pressure time series to calibrate the uncertain parameters of a popular subglacial drainage model, taking a Bayesian perspective to quantify the uncertainty in parameter estimates and in the calibrated model predictions. To reduce the computation time associated with Markov Chain Monte Carlo sampling, we construct a fast Gaussian process emulator to stand in for the subglacial drainage model. We first carry out a calibration experiment using synthetic observations consisting of model simulations with hidden parameter values as a demonstration of the method. Using real borehole water pressures measured in western Greenland, we find meaningful constraints on four of the eight model parameters and a factor-of-three reduction in uncertainty of the calibrated model predictions. These experiments illustrate Gaussian process-based Bayesian inference as a useful tool for calibration and uncertainty quantification of complex glaciological models using field data. However, significant differences between the calibrated model and the borehole data suggest that structural limitations of the model, rather than poorly constrained parameters or computational cost, remain the most important constraint on subglacial drainage modelling.

58 GEOSCIENCES↗

Revisiting the Architecture of the KOI-89 System

While high stellar obliquities observed in exoplanetary systems may be attributed to processes that tilt the planetary orbits, it is also possible that they reflect misalignments between protoplanetary disks and stellar spins. This latter hypothesis predicts the presence of coplanar multiplanetary systems misaligned with their central stars. Here we reevaluate the evidence of such an architecture that has been claimed for the KOI-89 system. An early-type star, KOI-89 has one validated transiting planet, KOI-89.01/Kepler-462b (period 84.7 days, radius 3.0 R {sub ⊕}), and one transiting planet candidate, KOI-89.02 (period 207.6 days, radius 4.0 R {sub ⊕}), where the latter exhibits transit timing variations (TTVs). A previous modeling of the stellar gravity-darkening effect in the transit light curves inferred a high stellar obliquity of ≈70°. We perform photodynamical modeling of the Kepler transit light curves and use the resulting constraints on the orbital configuration and transit times to update the gravity-darkened transit model. As a result, we find no firm evidence for the gravity-darkening effect in the transit shapes and conclude that stellar obliquity is not constrained by the data. Given the evidence for low orbital eccentricities from the dynamical analysis, the system architecture can thus be consistent with many other multitransiting systems with flat, near-circular orbits aligned with the stellar spin. We find that the TTVs imparted on its neighbor imply that KOI-89.01 has a mass ≳20 M {sub ⊕}. This would render it one of the densest known sub-Neptunes, mostly composed of a solid core. Lower masses are possible if the TTVs are instead due to an unseen third planet.

79 ASTRONOMY AND ASTROPHYSICS↗

Effect of Irradiation on Strengthening of a Model Fe-9Cr Oxide Dispersion Strengthened Alloy

The growing energy demand will increasingly call upon advanced nuclear reactors to supply safe and reliable energy worldwide. However, structural and cladding materials in these reactors will be subject to extreme conditions of up to a few hundred displacements per atom (dpa) at temperatures as high as 600 deg. C. Oxide dispersion strengthened (ODS) alloys are leading candidates for these components because their dispersion of Y-Ti-O nanoclusters provides high strength at elevated temperatures and dimensional stability under irradiation. But recent work has suggested that the oxides are unstable under irradiation, which has serious implications on mechanical performance. Thus, the objective of this work is to understand the change in strengthening mechanisms of an Fe-9Cr ODS alloy under irradiation. We use two approaches to assess strengthening of a proton-irradiated and neutron-irradiated Fe-9Cr ODS. First, nanoindentation directly measures changes in yield strength. Second, microstructure measurements are passed through models of two strengthening mechanisms. Results of these two approaches are compared to infer the strengthening mechanisms in the irradiated ODS alloy. A model Fe-9Cr ODS steel was irradiated to 3 dpa at 500 deg. C using either 2.0 MeV protons or a fast neutron spectrum. Both irradiation types induced the formation of dislocation loops and voids. Both proton and neutron irradiation resulted in dissolution of oxide nanoclusters, though neutron irradiation led to much more extensive oxide nano-cluster dissolution. The irradiated microstructure was used to predict irradiation hardening by considering both the precipitation strengthening and solid solution strengthening mechanisms. Due to more extensive oxide nano-cluster dissolution under neutron irradiation, the solid solution strengthening mechanism must be considered in order for the predicted change in yield strength to be comparable to that measured by nanoindentation. (authors)

36 MATERIALS SCIENCE↗

The behavior of transverse waves in nonuniform solar flux tubes. II. Implications for coronal loop seismology

The seismology of coronal loops using observations of damped transverse oscillations in combination with results from theoretical models is a tool to indirectly infer physical parameters in the solar atmospheric plasma. Existing seismology schemes based on approximations of the period and damping time of kink oscillations are often used beyond their theoretical range of applicability. These approximations assume that the variation of density across the loop is confined to a nonuniform layer much thinner than the radius of the loop, but the results of the inversion problem often do not satisfy this preliminary hypothesis. Here, we determine the accuracy of the analytic approximations of the period and damping time, and the impact on seismology estimates when largely nonuniform loops are considered. We find that the accuracy of the approximations when used beyond their range of applicability is strongly affected by the form of the density profile across the loop, that is observationally unknown and so must be arbitrarily imposed as part of the theoretical model. The error associated with the analytic approximations can be larger than 50% even for relatively thin nonuniform layers. This error directly affects the accuracy of approximate seismology estimates compared to actual numerical inversions. In addition, assuming different density profiles can produce noncoincident intervals of the seismic variables in inversions of the same event. The ignorance about the true shape of density variation across the loop is an important source of error that may dispute the reliability of parameters seismically inferred assuming an ad hoc density profile.

79 ASTRONOMY AND ASTROPHYSICS↗

Dark matter voids in the SDSS galaxy survey

What do we know about voids in the dark matter distribution given the Sloan Digital Sky Survey (SDSS) and assuming the ΛCDM model? Recent application of the Bayesian inference algorithm BORG to the SDSS Data Release 7 main galaxy sample has generated detailed Eulerian and Lagrangian representations of the large-scale structure as well as the possibility to accurately quantify corresponding uncertainties. Building upon these results, we present constrained catalogs of voids in the Sloan volume, aiming at a physical representation of dark matter underdensities and at the alleviation of the problems due to sparsity and biasing on galaxy void catalogs. To do so, we generate data-constrained reconstructions of the presently observed large-scale structure using a fully non-linear gravitational model. We then find and analyze void candidates using the VIDE toolkit. Our methodology therefore predicts the properties of voids based on fusing prior information from simulations and data constraints. For usual void statistics (number function, ellipticity distribution and radial density profile), all the results obtained are in agreement with dark matter simulations. Our dark matter void candidates probe a deeper void hierarchy than voids directly based on the observed galaxies alone. The use of our catalogs therefore opens the way to high-precision void cosmology at the level of the dark matter field. We will make the void catalogs used in this work available at http://www.cosmicvoids.net.

79 ASTRONOMY AND ASTROPHYSICS↗

ABUNDANCE OF {sup 26}Al AND {sup 60}Fe IN EVOLVING GIANT MOLECULAR CLOUDS

The nucleosynthesis and ejection of radioactive {sup 26}Al (t{sub 1/2} {approx} 0.72 Myr) and {sup 60}Fe, (t{sub 1/2} {approx} 2.5 Myr) into the interstellar medium is dominated by the stellar winds of massive stars and supernova type II explosions. Studies of meteorites and their components indicate that the initial abundances of these short-lived radionuclides in the solar protoplanetary disk were higher than the background levels of the galaxy inferred from {gamma}-ray astronomy and models of the galactic chemical evolution. This observation has been used to argue for a late-stage addition of stellar debris to the solar system's parental molecular cloud or, alternatively, the solar protoplanetary disk, thereby requiring a special scenario for the formation of our solar system. Here, we use supercomputers to model-from first principles-the production, transport, and admixing of freshly synthesized {sup 26}Al and {sup 60}Fe in star-forming regions within giant molecular clouds. Under typical star formation conditions, the levels of {sup 26}Al in most star-forming regions are comparable to that deduced from meteorites, suggesting that the presence of short-lived radionuclides in the early solar system is a generic feature of the chemical evolution of giant molecular clouds. The {sup 60}Fe/{sup 26}Al yield ratio of Almost-Equal-To 0.2 calculated from our simulations is consistent with the galactic value of 0.15 {+-} 0.06 inferred from {gamma}-ray astronomy but is significantly higher than most current solar system measurements indicate. We suggest that estimates based on differentiated meteorites and some chondritic components may not be representative of the initial {sup 60}Fe abundance of the bulk solar system.

79 ASTRONOMY AND ASTROPHYSICS↗

Generalized framework for likelihood-based field-level inference of growth rate from velocity and density fields

Measuring the growth rate of large-scale structures ( f ) as a function of redshift has the potential to break degeneracies between modified gravity and dark energy models, when combined with expansion-rate probes. Direct estimates of peculiar velocities of galaxies have attracted interest as a means of estimating fσ 8 . In particular, field-level methods can be used to fit the field nuisance parameter along with cosmological parameters simultaneously. This article aims to provide the community with a unified framework for the theoretical modeling of the likelihood-based field-level inference by performing fast field covariance calculations for velocity and density fields. Our purpose is to lay the foundations for a nonlinear extension of the likelihood-based method at the field level. We have developed a generalized framework, implemented in the dedicated software flip to perform a likelihood-based inference of fσ 8 . We derived a new field covariance model, which includes wide-angle corrections. We also included the models previously described in the literature inside our framework. We compared their performance against ours, and we validated our model by comparing it with the two-point statistics of a recent N-body simulation. The tests we performed have allowed us to validate our software and determine the appropriate wavenumber range to integrate our covariance model and its validity in terms of separation. Our framework allows for a wider wavenumber coverage to be used in our calculations than in previous works, which is particularly interesting for nonlinear model extensions. Finally, our generalized framework allows us to efficiently perform a survey geometry-dependent Fisher forecast of the fσ 8 parameter. We show that the Fisher forecast method we developed gives an error bar that is 30% closer to a full likelihood-based estimation than a standard volume Fisher forecast.

Ravoux, Corentin↗

Hybrid Star Models in the Light of New Multimessenger Data

Abstract Recent astrophysical mass inferences of compact stars HESS J1731-347 and PSR J0952-0607, with extremely small and large masses respectively, as well as the measurement of the neutron skin of Ca in the CREX experiment challenge and constrain the models of dense matter. We examine the concept of hybrid stars—objects containing quark cores surrounded by nucleonic envelopes—as models that account for these new data along with other inferences. We employ a family of 81 nucleonic equations of state (EOSs) with variable skewness and slope of symmetry energy at saturation density and a constant speed-of-sound EOS for quark matter. For each nucleonic EOS, a family of hybrid EOSs is generated by varying the transition density, the energy jump, and the speed of sound. These models are tested against the data from GW170817 and J1731-347, which favor low-density soft EOS and J0592-0607 and J0740+6620, which require high-density stiff EOS. The addition of J0592-0607's mass measurement to the constraints has no significant impact on the parameter space of the admissible EOS, but allows us to explore the potential effect of pulsars more massive than J0740+6620, if such exists. We then examine the occurrence of twin configurations and quantify the ranges of masses and radii that they can possess. It is shown that including J1731-347 data favors EOSs that predict low-mass twins with M ≲ 1.3 M ⊙ that can be realized if the deconfinement transition density is low. If combined with large speed of sound in quark matter such models allow for maximum masses of hybrid stars in 2.0–2.6 M ⊙ .

Astronomy & Astrophysics↗

Quantification and prediction of solidification textures under additive manufacturing conditions

Crystallographic textures are a major determinant of the macroscale anisotropic properties of polycrystalline metallic alloys produced in a wide range of additive manufacturing (AM) processes. Here, we introduce a statistical method that can accurately quantify the degree of orientational order of textures despite the large random fluctuations in the orientation of individual grains inherent in AM processes. The method, demonstrated for laser and resolidification of AlSi thin films, extends Z-scoring to a dynamical regime to assess the statistical significance of observed textures compared to randomly generated ones at different stages of solidification. We further show that, combined with phase-field modeling, this method can be used to infer fundamental anisotropic properties of the solid-liquid interface that are essential for texture prediction, and are compared here to the results of atomistic simulations. In addition, phase-field modeling reveals that, even at rapid AM solidification rates, the observed 〈110〉-dominated textures in the AlSi thin films are controlled predominantly by the anisotropy of the interface free-energy and sheds light on the physical mechanism of grain competition. These results significantly enhance both the existing tools for the quantification and prediction of AM crystallographic textures and our basic understanding of their formation.

36 MATERIALS SCIENCE↗

Interpreting and Accelerating Transformers for Jet Tagging

Attention-based transformers are ubiquitous in machine learning applications from natural language processing to computer vision. In high energy physics, one central application is to classify collimated particle showers in colliders based on the particle of origin, known as jet tagging. In this work, we study the interpretatbility and prospects for acceleration of Particle Transformer (ParT), a state-of-the-art model, leverages particle-level attention to improve jet-tagging performance. We analyzing ParT's attention maps and particle-pair correlations in the eta-phi plane, revealing intriguing features, such as a binary attention pattern that identifies critical substructure in jets. These insights enhance our understanding of the model's internal workings and learning process and hint at ways to improve its efficiency. Along these lines, we also explore low-rank attention, attention alternatives, and dynamic quantization to accelerate transformers for jet tagging. With quantization, we achieve a 50% reduction in model size and a 10% increase in inference speed without compromising accuracy. These combined efforts enhance both the performance and the interpretability of transformers in high-energy physics, opening avenues for more efficient and physics-driven model designs.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗

Kepler’S supernova: An OVERLUMINOUS type Ia event interacting with a massive circumstellar medium at a very late phase

We have analyzed XMM-Newton, Chandra, and Suzaku observations of three similarly aged, ejecta-dominated supernova remnants (SNRs), Kepler's SNR, Tycho's SNR, and SNR 0509-67.5, to investigate the properties of the SN ejecta and the circumstellar medium (CSM). By simply comparing the X-ray spectra, we find that line intensity ratios of iron-group elements (IGEs) to intermediate-mass elements (IMEs) for Kepler's SNR and SNR 0509-67.5 are much higher than those for Tycho's SNR. Given that Tycho's SNR and SNR 0509-67.5 are thought to be the remnants of a typical SN Ia and an overluminous SN Ia, respectively, we argue that Kepler is the product of an overlumious SN Ia. This inference is supported by our spectral modeling, which reveals the IGE and IME masses, respectively, to be 0.95{sub −0.37}{sup +0.34} M{sub ⊙} and 0.12{sub −0.05}{sup +0.19} M{sub ⊙} (Kepler's SNR), 0.75{sub −0.15}{sup +0.51} M{sub ⊙} and 0.34{sub −0.25}{sup +0.08} M{sub ⊙} (SNR 0509-67.5), and 0.35{sub −0.15}{sup +0.55} M{sub ⊙} and 0.70{sub −0.28}{sup +0.12} M{sub ⊙} (Tycho's SNR). We find that the CSM component in Kepler's SNR consists of tenuous diffuse gas (∼0.3 M{sub ⊙}) present throughout the entire remnant, plus dense knots (∼0.035 M{sub ⊙}). Since both of them show N overabundance, their origin would be CNO-processed material from the progenitor system. The mass of the diffuse CSM allows us to infer the pre-SN mass-loss rate to be ∼1.5×10{sup −5}(v{sub w}/10 km s{sup −1}) M{sub ⊙} yr{sup −1}. The dense knots have slow proper motions and relatively small ionization timescales; hence, they were likely located a few parsecs away from the progenitor at the explosion. We thus argue that Kepler's SN was an overluminous (91T-like) event that recently started to interact with the massive CSM. This supports the possible link between 91T-like SNe and “Ia-CSM” SNe.

79 ASTRONOMY AND ASTROPHYSICS↗

The effects of viewing angle on the inference of magnetic shear in preflare active regions

Analytic models for nonpotential force-free fields representative of preflaring active regions have been used to determine the degree of magnetic shear along the magnetic neutral line of these fields as a function of location and orientation of the active region on the solar disk. The results indicate that, with the exception of regions close to the disk center, the position of the inferred neutral line differs significantly from the actual neutral line. It is suggested that the significant variation of the inferred degree of shear with the position and orientation of the region is due to such factors as geometric projection effects, the shift of the inferred neutral line relative to its true position, and the variations in the reference potential field.

Wilkinson, Ladye Kathryn↗

MEAN AGE GRADIENT AND ASYMMETRY IN THE STAR FORMATION HISTORY OF THE SMALL MAGELLANIC CLOUD

We derive the star formation history (SFH) in four regions of the Small Magellanic Cloud (SMC) using the deepest VI color-magnitude diagrams (CMDs) ever obtained for this galaxy. The images were obtained with the Advanced Camera for Surveys on board the Hubble Space Telescope (HST) and are located at projected distances of 0.°5-2° from the SMC center, probing the main body and the wing of the galaxy. We derived the SFHs of the four fields using two independent procedures to fit synthetic CMDs to the data. We compare the SFHs derived here with our earlier results for the SMC bar to create a deep pencil-beam survey of the global history of the central SMC. We find in all the six fields observed with HST a slow star formation (SF) pace from 13 to 5-7 Gyr ago, followed by a ≈2-3 times higher activity. This is remarkable because dynamical models do not predict a strong influence of either the Large Magellanic Cloud (LMC) or the Milky Way at that time. The level of the intermediate-age SF rate enhancement systematically increases toward the center, resulting in a gradient in the mean age of the population, with the bar fields being systematically younger than the outer ones. SF over the most recent 500 Myr is strongly concentrated in the bar, the only exception being the area of the SMC wing. The strong current activity of the latter is likely driven by interaction with the LMC. At a given age, there is no significant difference in metallicity between the inner and outer fields, implying that metals are well mixed throughout the SMC. The age-metallicity relations we infer from our best-fitting models are monotonically increasing with time, with no evidence of dips. This may argue against the major merger scenario proposed by Tsujimoto and Bekki in 2009, although a minor merger cannot be ruled out.

79 ASTRONOMY AND ASTROPHYSICS↗

Understanding protein-complex assembly through grand canonical maximum entropy modeling

Inside a cell, heterotypic proteins assemble in inhomogeneous, crowded systems where the abundance of these proteins vary with cell types. While some protein complexes form putative structures that can be visualized with imaging, there are far more protein complexes that are yet to be solved because of their dynamic associations with one another. Nevertheless, it is possible to infer these protein complexes through a physical model. However, it is often not clear to physicists what kind of data from biology is necessary for such a modeling endeavor. Here, we aim to model these clusters of coarse-grained protein assemblies from multiple subunits through the constraints of interactions among the subunits and the chemical potential of each subunit. We obtained the constraints on the interactions among subunits from the known protein structures. We inferred the chemical potential that dictates the particle number distribution of each protein subunit from the knowledge of protein abundance from experimental data. Guided by the maximum entropy principle, we formulated an inverse statistical mechanical method to infer the distribution of particle numbers from the data of protein abundance as chemical potentials for a grand canonical multicomponent mixture. Using grand canonical Monte Carlo simulations, we captured a distribution of high-order clusters in a protein complex of succinate dehydrogenase with four known subunits. The complexity of hierarchical clusters varies with the relative protein abundance of each subunit in distinctive cell types such as lung, heart, and brain. When the crowding content increases, we observed that crowding stabilizes emergent clusters that do not exist in dilute conditions. We, therefore, proposed a testable hypothesis that the hierarchical complexity of protein clusters on a molecular scale is a plausible biomarker of predicting the phenotypes of a cell.

59 BASIC BIOLOGICAL SCIENCES↗

INTEGRAL FIELD SPECTROSCOPY AND MULTI-WAVELENGTH IMAGING OF THE NEARBY SPIRAL GALAXY NGC 5668 : AN UNUSUAL FLATTENING IN METALLICITY GRADIENT

We present an analysis of the full bidimensional optical spectral cube of the nearby spiral galaxy NGC 5668, observed with the Pmas fiber PAcK Integral Field Unit (IFU) at the Calar Alto observatory 3.5 m telescope. We make use of broadband imaging to provide further constraints on the evolutionary history of the galaxy. This data set will allow us to improve our understanding of the mechanisms that drive the evolution of disks. We investigated the properties of 62 H II regions and concentric rings in NGC 5668 and derived maps in ionized-gas attenuation and chemical (oxygen) abundances. We find that while inward of r {approx}36'' {approx} 4.4 kpc {approx} 0.36 (D{sub 25}/2) the derived O/H ratio follows the radial gradient typical of spiral galaxies, the abundance gradient beyond r {approx} 36'' flattens out. The analysis of the multi-wavelength surface brightness profiles of NGC 5668 is performed by fitting these profiles with those predicted by chemo-spectrophotometric evolutionary models of galaxy disks. From this, we infer a spin and circular velocity of {lambda} = 0.053 and v{sub c} = 167 km s{sup -1}, respectively. The metallicity gradient and rotation curve predicted by this best-fitting galaxy model nicely match the values derived from the IFU observations, especially within r {approx}36''. The same is true for the colors despite some small offsets and a reddening in the bluest colors beyond that radius. On the other hand, deviations of some of these properties in the outer disk indicate that a secondary mechanism, possibly gas transfer induced by the presence of a young bar, must have played a role in shaping the recent chemical and star formation histories of NGC 5668.

79 ASTRONOMY AND ASTROPHYSICS↗

GCoD: Graph Convolutional Network Acceleration via Dedicated Algorithm and Accelerator Co-Design

Graph Convolutional Networks (GCNs) have emerged as the state-of-the-art graph learning model. However, it remains notoriously challenging to inference GCNs over large graph datasets, limiting their application to large real-world graphs and hindering the exploration of deeper and more sophisticated GCN graphs. This is because real-world graphs can be extremely large and sparse. Furthermore, the node degree of GCNs tends to follow the power-law distribution and therefore have highly irregular adjacency matrices, resulting in prohibitive inefficiencies in both data processing and movement and thus substantially limiting the achievable GCN acceleration efficiency. To this end, this paper proposes the first GCN algorithm and accelerator Co-Design framework dubbed GCoD which can largely alleviate the aforementioned GCN irregularity and boost GCNs' inference efficiency. Specifically, on the algorithm level, GCoD integrates a divide and conquer GCN training strategy that polarizes the graphs to be either denser or sparser in local neighborhoods without compromising the model accuracy, resulting in graph adjacency matrices that (mostly) have merely two levels of workload and enjoys largely enhanced regularity and thus ease of acceleration. On the hardware level, we further develop a dedicated two-pronged accelerator with a separated engine to process each of the aforementioned workloads, further boosting the overall utilization and acceleration efficiency. Extensive experiments and ablation studies validate that our GCoD consistently outperforms state-of-the-art designs in terms of accelerator efficiency while maintaining or even improving the task accuracy. Additionally, we visualize GCoD trained graph adjacency matrices to better understand its advantages. All codes and pre-trained models will be released upon acceptance.

You, Haoran↗